humy65/hebrew-intent-classifier

Hebrew Intent Classification Model

Model Description

This model is a fine-tuned BERT model for Hebrew intent classification, specifically designed for customer service scenarios. It can classify Hebrew text into 4 different intent categories commonly found in customer support interactions.

Supported Intent Classes

1. ביטול מנוי (Cancel Subscription) - Requests to cancel or terminate services
2. שאלה כללית (General Question) - General inquiries about services, pricing, or account management
3. שכחת סיסמה (Password Reset) - Issues related to forgotten passwords or login problems
4. תמיכה טכנית (Technical Support) - Technical issues, bugs, or system problems

Usage

from transformers import pipeline

# Load the model
classifier = pipeline("text-classification", model="Huggingm1r@n/hebrew-intent-classifier")

# Make predictions
result = classifier("שכחתי את הסיסמה שלי")
print(result)
# [{'label': 'שכחת סיסמה', 'score': 0.95}]

# Test other examples
examples = [
    "רוצה לבטל את המנוי",
    "כמה עולה החבילה", 
    "האתר לא עובד"
]

for text in examples:
    result = classifier(text)
    print(f"'{text}' -> {result[0]['label']} ({result[0]['score']:.2%})")

Direct Usage with Transformers

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

# Load model and tokenizer
tokenizer = AutoTokenizer.from_pretrained("Huggingm1r@n/hebrew-intent-classifier")
model = AutoModelForSequenceClassification.from_pretrained("Huggingm1r@n/hebrew-intent-classifier")

def predict_intent(text):
    inputs = tokenizer(text, return_tensors="pt", padding=True, truncation=True)
    
    with torch.no_grad():
        outputs = model(**inputs)
        logits = outputs.logits
        probabilities = torch.softmax(logits, dim=-1)
    
    predicted_id = torch.argmax(logits, dim=-1).item()
    predicted_label = model.config.id2label[predicted_id]
    confidence = probabilities[0][predicted_id].item()
    
    return predicted_label, confidence

# Example
intent, confidence = predict_intent("שכחתי את הסיסמה")
print(f"Intent: {intent}, Confidence: {confidence:.2%}")

Training Details

• Base Model: bert-base-multilingual-cased
• Training Data: 135 Hebrew customer service examples (augmented from 12 original)
• Data Augmentation: Manual variations, formal/informal styles, polite forms
• Performance: >90% accuracy on validation set

Example Predictions

Hebrew Text	Predicted Intent	English Translation
שכחתי את הסיסמה שלי	שכחת סיסמה	I forgot my password
רוצה לבטל את המנוי	ביטול מנוי	Want to cancel subscription
כמה עולה החבילה	שאלה כללית	How much does the package cost
האתר לא עובד	תמיכה טכנית	The website doesn't work

Use Cases

• Customer Service Chatbots: Route Hebrew customer queries automatically
• Support Ticket Classification: Categorize support requests by intent
• Voice of Customer Analysis: Analyze Hebrew customer feedback
• Automated Response Systems: Trigger appropriate responses based on intent

Limitations

• Designed for customer service domain specifically
• Limited to 4 predefined intent classes
• May not work well with very informal Hebrew or slang
• Requires Hebrew text input

Model Files

• Uses safetensors format for secure model storage
• Compatible with latest Transformers library
• Includes comprehensive tokenizer configuration

Citation

@misc{hebrew-intent-classifier-2025,
  title={Hebrew Intent Classification Model for Customer Service},
  author={Huggingm1r@n},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/Huggingm1r@n/hebrew-intent-classifier}
}

License

This model is released under the Apache 2.0 License.